Characterization of Wood Materials Using Perception-Related Image Statistics

An efficient computational characterization of real-world materials is one of the challenges in image understanding. An automatic assessment of materials, with similar performance as human observer, usually relies on complicated image filtering derived from models of human perception. However, these models become too complicated when a real material is observed in the form of dynamic stimuli. This study tackles the challenge from the other side. First, we collected human ratings of the most common visual attributes for videos of wood samples and analyzed their relationship to selected image statistics. In our experiments on a set of sixty wood samples, we have found that such image statistics can perform surprisingly well in the discrimination of individual samples with reasonable correlation to human ratings. We have also shown that these statistics can be also effective in the discrimination of images of the same material taken under different illumination and viewing conditions.